Optimal scheduling of renewable power distribution systems combining deep learning and particle swarm optimization

Introduction: To address the optimal scheduling problem of renewable power distribution systems, this paper proposes an integrated framework combining deep learning-based load forecasting with particle swarm optimization. Objectives: The first objective is accurate multivariate load forecasting (cooling, heating, and electric loads). The second objective is minimizing network loss by optimally scheduling electric vehicle (EV) charging locations and times. Methods: For load forecasting, a hybrid RCNN-SVR model is constructed. The convolutional neural network (CNN) acts as a feature extractor to implicitly capture representative patterns from input data, while support vector regression (SVR) produces the final load predictions. Missing and outlier data are pre-processed. For optimal scheduling, a dual-layer particle swarm optimization (PSO) algorithm is developed. The inner layer enforces system constraints, and the outer layer minimizes network loss. EV charging load is simulated using the Monte Carlo method, and two cases (variable vs. fixed charging addresses) are optimized. Results: Experimental results demonstrate that the proposed RCNN-SVR model achieves high prediction accuracy, with mean absolute percentage error as low as 2.41% for winter electric loads. The dual-layer PSO reduces peak system load from 11.2×10³ kW to 10.5×10³ kW and decreases network loss by 2.13%, effectively smoothing grid fluctuations. Conclusion: The RCNN-SVR model significantly improves multivariate load prediction accuracy compared to separate forecasting methods. The dual-layer PSO successfully converts disorderly EV charging into orderly scheduling, reducing peak load and network loss. Together, they provide a practical solution for renewable distribution system scheduling.